Mechanical Systems and Signal Processing 94 (2017) 111128

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Speed tracking control of pneumatic motor servo systems using

observation-based adaptive dynamic sliding-mode control
Syuan-Yi Chen a,, Sheng-Sian Gong b
a
Department of Electrical Engineering, National Taiwan Normal University, Taipei 106, Taiwan
b
Display System Products Group, Qisda Corporation, Taoyuan, 333, Taiwan

a r t i c l e i n f o a b s t r a c t

Article history: This study aims to develop an adaptive high-precision control system for controlling the
Received 1 November 2016 speed of a vane-type air motor (VAM) pneumatic servo system. In practice, the rotor speed
Received in revised form 9 January 2017 of a VAM depends on the input mass air flow, which can be controlled by the effective ori-
Accepted 16 February 2017
fice area (EOA) of an electronic throttle valve (ETV). As the control variable of a second-
order pneumatic system is the integral of the EOA, an observation-based adaptive dynamic
sliding-mode control (ADSMC) system is proposed to derive the differential of the control
Keywords:
variable, namely, the EOA control signal. In the ADSMC system, a proportionalintegral
Fuzzy neural network
Pneumatic servo system
derivative fuzzy neural network (PIDFNN) observer is used to achieve an ideal dynamic
Speed control sliding-mode control (DSMC), and a supervisor compensator is designed to eliminate the
Dynamic sliding-mode control approximation error. As a result, the ADSMC incorporates the robustness of a DSMC and
Vane-type air motor the online learning ability of a PIDFNN. To ensure the convergence of the tracking error,
a Lyapunov-based analytical method is employed to obtain the adaptive algorithms
required to tune the control parameters of the online ADSMC system. Finally, our experi-
mental results demonstrate the precision and robustness of the ADSMC system for highly
nonlinear and time-varying VAM pneumatic servo systems.
2017 Elsevier Ltd. All rights reserved.

1. Introduction

Renewable energy sources have recently attracted a lot of attention because of increasing environmental concerns and
stringent laws for regulating the use of non-renewable energy sources. As a result, compressed air energy has found appli-
cations in a number of industrial areas such as manufacturing, transportation, automation, and bionics because of the advan-
tages of this energy source such as zero pollutant-emission, simple structure, and outstanding control performance [1]. To
convert pneumatic energy into mechanical motion, various pneumatic drives have been used in different areas such as actu-
ator systems [3,4], servo systems [5,6], locomotives brake [7], manipulator [8], beams system [9], and artificial muscles [10].
Unlike electric or hydraulic actuators, servo pneumatic actuators are characterized by their clear usage, simple maintenance,
easy implementation, low cost, and high power density. In addition, servo pneumatic actuators are able to reduce the pos-
sibility of electrical shocks or fires occurring in high temperature, flammable, or explosive operational environments. The
demand for high-accuracy air motor systems that can be used in industrial applications has increased in recent years,
and numerous scientists have started using servo-controlled pneumatic air motors to develop complicated motion control
tasks [1116].

Corresponding author at: No. 129, He-ping E. Rd. Sec. 1, Taipei 10610, Taiwan.
E-mail address: chensy@ntnu.edu.tw (S.-Y. Chen).

http://dx.doi.org/10.1016/j.ymssp.2017.02.025
0888-3270/ 2017 Elsevier Ltd. All rights reserved.
112 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128

System modeling and dynamic analysis are crucial for accurately predicting system behavior and efficiently designing
control strategies. In one study, linear and rotary pneumatic actuators were developed [3], and two pneumatic circuit for-
mulas, i.e., an air thermodynamic transformation equation and a corresponding energy equation, were derived using a bond
graph approach. Other studies have proposed using scroll air motors as pneumatic actuators [17,18]. These studies reported
a model developed on the basis of adiabatic processes and perfect gas assumptions. A control-oriented vane-type air motor
(VAM) model was developed in another study and it used two basic elements; an isentropic nozzle and a control volume
[19]. In this study, parameters such as temperature and pressure inside the chamber of the motor, speed and torque of
the motor, and the mass air flow (MAF) rate of air consumption were simulated and experimentally verified in a detailed
manner.
Pneumatic servo systems demonstrate highly nonlinear and time-varying behaviors because of the compressibility of air,
friction of mechanical components, dead-zone effect of control valves, eccentric mechanisms, and aerodynamic forces [36].
These systems will yield different responses even with the same inlet MAF [13]. As such, the nonlinear characteristics of
these systems make their control as challenging issues. A study developed and experimentally evaluated a proportionalin-
tegral (PI) position controller that can be used in a low-cost industrial pneumatic actuator [4]. A function approximation
technique based adaptive controller has also been proposed for use in pneumatic servo systems with variable payload
and uncertain disturbances [5]. Another study proposed using a state-feedback nonlinear controller for a pneumatic cylinder
based on the theory of homogeneous, finite-time stable, ordinary differential equations [6]. One study proposed using a
sliding-mode control (SMC) system with a non-linear disturbance observer to control a pneumatic muscle actuator that
can be used in a hand rehabilitation device [10]. Another study considered using a neural network (NN)-based model-
free controller to achieve high-accuracy speed control in a VAM [16]. In the present study, a novel adaptive control strategy,
different from these aforementioned methods, has been developed.
Proportionalintegralderivative (PID) controllers are widely used in industries and academia. However, conventional
PID controllers do not exhibit satisfactory control over highly nonlinear and time-varying systems because of their linear
structures with constant control gains. To deal with this problem, various approaches were proposed to improve the control
performances of the conventional PID or PI controllers. A general type-2 fuzzy logic sets and a modified harmony search
algorithm were integrated to tune the proportional and integral gains of the PI controller adaptively [20]. Moreover, PID neu-
ral networks (PIDNN) [21] have been developed to tune the proportional, integral, and derivative gains of a PID controller
online. Therefore, the PIDNN reinforces the traditional PID control with adaptiveness. On the other hand, a proportionalinte
gralderivative fuzzy neural network (PIDFNN), which builds upon the advantageous characteristics of the PIDNN and fuzzy
neural network (FNN) [22], has been proposed [16]. Unlike conventional FNNs, the PIDFNN uses an auxiliary multi-input
multi-output (MIMO) PIDNN within its architecture to perform adaptive PID control with online tuning fuzzy rules. Exper-
imental results for the PIDFNN demonstrated better control and learning abilities than that demonstrated by traditional PID
and PIDNN controllers.
SMC is a widely studied control approach that provides robustness against certain disturbances and system uncertainties
[2330]. However, the required switching function for the hitting control results in high-frequency chattering and leads to
undesirable results such as low control accuracy and high wear of moving mechanical parts. Therefore, a variety of solutions
have been proposed to suppress the chattering phenomenon [2328]. Among these methods, dynamic SMC (DSMC) is an
effective method for eliminating the chattering that has received a lot of attention in recent years [23]. Besides, the differ-
ential of the control signal can be obtained directly according to this design. However, the boundary for realizing the hitting
control of a DSMC is always vague and changeable in a practical control system. On the other hand, some studies applied
fuzzy approximation methods in the SMC in order to directly estimate the uncertainties to deal with the undesirable chat-
tering phenomenon [24,25]. Furthermore, an optimal type II fuzzy SMC approach was presented to not only warranties the
constancy and hardiness against uncertainties, but also considerably decreases the control chattering inherent in traditional
SMC [26]. Another adaptive fractional order PID sliding mode controller (AFOPIDSMC) using a Bat algorithm was proposed to
control a Caterpillar robot manipulator [27]. In the AFOPIDSMC, a new combined control law was developed for chattering
reduction by means of fractional order PID controller and high trajectory tracking through using SMC. Additionally, an adap-
tive neural network integral SMC was proposed to control a biped robot [28]. To eliminate the chattering phenomena, an
adaptive neural network was used to estimate the unknown disturbances of the robot directly.
In general, the control methods for the VAM type pneumatic servo systems can be majorly classified into two classes
regarding the ways for deriving a control signal. The first class is the model-free control, such as fuzzy control [11] and
PID control [15], in which the control systems are used to derive the control signals without knowing the exact model of
the system. According to the control of the effective orifice area (EOA) of the electronic throttle valve (ETV), the speed of
the VAM can be adjusted well under some specific operating conditions. Nevertheless, favorable control performance levels
are guaranteed only under these conditions because the control strategies or algorithms encounter difficulties in considering
all the operating situations in advance. Moreover, the robustness of the control systems are weak because the parameters of
the controller are not adaptive. The second class is the model-based control such as backstepping control [13] and SMC [14].
These control systems assumed that the difference between the upstream and downstream pressures of the VAM is propor-
tional to the EOA. Thus, the control systems are used to control the pressure difference for regulating the output torques
directly, thereby changing the rotor speed. However, in practice, the EOA possesses a complicated one-order dynamic with
respect to the pressure [2], rather than the simple proportional relation. Therefore, the control performance levels are
restricted due to the omission of this one-order dynamic. Furthermore, considering the highly nonlinear and time-varying
 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128 113

behaviors of the working air in the chamber and the friction in the mechanical system, deriving the system parameters for
implementing the control laws of such systems are extremely difficult. The inaccuracies of the system parameters may result
in steady-state error and poor robustness in practical VAM applications.
For addressing the aforementioned drawbacks, the objective of this study is to develop a new adaptive control scheme
based on a more realistic model. Firstly, the operating principle and configuration of the VAM pneumatic servo system
are presented. Subsequently, a new second-order dynamic model that can appropriately express the relation between the
EOA of the ETV and the consequent torque of the VAM, according to a basic thermal dynamic analysis is presented. Because
only the EOA is controllable in a practical pneumatic servo system, a DSMC system is designed to derive the differential of
the control variable so that the direct control of the EOA becomes feasible. However, because uncertainties are difficult to
measure, an observation-based adaptive dynamic sliding-mode control (ADSMC) system is further proposed to imitate an
ideal DSMC system without uncertainties information. Finally, experimental results demonstrating the effectiveness of
the proposed ADSMC system for the VAM pneumatic servo system are discussed and compared. The main findings of this
study are summarized as follows:

(1) Modeling of the VAM pneumatic servo system: The developed second-order model of the VAM concerns the practical
relation between the EOA and the resultant torque, and it is significant for featuring the dynamics of the VAM.
(2) Development and application of the ADSMC system: The proposed ADSMC system with adaptive laws incorporates the
robustness of a DSMC and the online learning ability of a PIDFNN and is used to control the speed of the VAM directly.
(3) Comparison of two cases between five controllers: Experiments regarding nominal and payload cases were under-
taken to compare the performance levels of PID, PIDNN, FNN, PIDFNN, and ADSMC controllers. The results show that
the proposed ADSMC system exhibits favorable speed control performance with regard to the VAM pneumatic servo
system.

2. System configuration and modeling

2.1. Operating principle of VAM

A VAM with four vanes (GAST 4AM-FRV-13C, output power of 1.7 hp, maximum pressure of 7 bar, and maximum speed of
3000 rpm), as shown in Fig. 1(a), was used as the power source for transferring the pneumatic energy into rotational
mechanical force. A cylindrical frame forms the stator shell outermost the VAM. A rotor is eccentric to the center of the stator
shell and has four mutually perpendicular guide channels, each of which is fitted with a freely sliding rectangular vane.
When the vanes rotate with the rotor, they passively tend to either slide out in the channels because of the centrifugal force
or move back into the channels because of the mechanical structure. Moreover, two neighboring vanes form a chamber space
in the stator. Thus, the four chambers have dissimilar instant volumes because of the eccentric feature between the stator
and rotor. Depending on the MAF direction, the motor rotates in a counterclockwise direction, as shown in Fig. 1(a).
Regarding the operation of the VAM, the compressed MAF enters the first chamber through an intake port. The momen-
tum of the MAF then acts as the mechanical force acting on the surface of the first vane, as shown in Fig. 1(a). It forms the
rotational torque and accelerates the rotational speed of the rotor. When the rotor rotates counterclockwise, the second
chamber moves from the exhaust-air part to the intake-air part. Accordingly, the air flowing out of the exhaust port is refilled
again. In a similar manner, the pneumatic power continuously provides rotational torque to the rotor. The third and fourth
chambers work according to the same mechanism. Fig. 1(b) shows the different motor speeds caused by different air con-
sumptions under different supplied air pressures in the VAM [31].

2.2. Configuration of the VAM pneumatic servo system

Fig. 2 illustrates the complete configuration of the VAM pneumatic servo system. First, the air is compressed by a com-
pressor. Then, the compressed air is dried to reduce the quantity of water and hydrocarbon vapors to improve the process
efficiency. Subsequently, a high-pressure air tank acts as the energy source, as it stores the compressed air used to propel the
rotor of the VAM. The pressure of the output air is limited by a pressure regulator downstream of the output port of the tank.
Two air pipes with different diameters connect the pressure regulator and the ETV composed of a piston actuated valve
(BPG206STX23) and an electropneumatic transducer (Type500-EH). They are connected by thick and thin pipes, respectively,
as shown in Fig. 2. A filter between the pressure regulator and the electropneumatic transducer is used to avoid the contam-
inants from entering the transducer. The output air pressure of the electropneumatic transducer can be linearly controlled by
an external applied voltage signal. The linear ranges of the input voltage and output air pressure are 010 V and 3120 PSIG,
respectively. After that, the output compressed air with controlled pressures acts on a normal closed piston actuated valve so
as to gradually open the valve with a spring return action. In other words, the EOA of the valve can be proportionally adjusted
by an external voltage signal applied to the electropneumatic transducer. Thus, the MAF rate flowing into the VAM through
the EOA of the ETV can regulate the speed well.
For practical implementation, the voltage applied to the ETV in the range of 210 V can achieve 0100% EOA linearly. In
this study, a digital signal processor (DSP, TMS320F28335)-based VAM controller receives the rotor angle and speed
114 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128

Input Air
0 / 360 Exhausted Air
ETV

Intake Port Exhaust

E Port
1st Chamberr 2nd Chamber
Rotor
90 270

Vane
V
Stator Shell

4th Chamber
b 3rd Chamber

180
a
35 7.0 bar
30
5.6 bar
Air Consumption (l/s)

25
4.2 bar
20
2.8 bar
15

10 1.4bar

5
0
0 500 1000 1500 2000 2500 3000
Speed (rpm)
b
Fig. 1. Structure and characteristics of a VAM: (a) structure and operating principle of a VAM, (b) different rotor speeds caused by different volume air flow
consumptions at various pressure settings.

information from an encoder (HTR-HM-15-500-3-L) to determine the control signal for controlling the EOA. A magnetic
powder brake (POC-200A) is also used to simulate the external load in the VAM pneumatic servo system.
In summary, the MAF starts from the air tank, and then moves through the pressure regulator. After this, it moves through
the ETV before, as a final step, entering the input port of the VAM, where it moves the vanes inside the motor. The total air
force collected on the four vanes then provides torque to the rotor. This torque is then applied against the rotor inertia and
external load. The rotor accelerates or decelerates depending on the net torque of the traction and load torques.

2.3. Dynamic analysis of VAM

The MAF rate of a pneumatic servo system that passes through the EOA of the ETV can be described as follows [4,19]:
sq  c1
c
C d Pu / 2c c1 Pd 2
_ p
Unchoked flow : m Pd =Pu 2=c  P d =Pu c ; > 1
Tu Rc  1 Pu c1

 c1=2c1  c1
c
C Pu / 1=2 2c Pd 2
_ pd
Choked flow : m c ; 6 ; 2
RT u c1 Pu c1
where C d , R, and c represent the discharge factor, ideal gas constant, and specific heat ratio, respectively; Pu ; Pd ; and T u are
the upstream pressure, downstream pressure, and upstream temperature, respectively; and / is the EOA of the ETV, which
can be controlled via an externally applied voltage. Eq. (2) denotes that once the upstream/downstream pressure difference
is increased to a certain value, the MAF rate will no longer be affected by the pressure difference. According to a basic
 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128 115

Compressor Pressure Piston Actuated Magnetic

Regulator Valve VAM Powder Brake

EOA

Filter
Encoder

Electropneumatic Personal
Compressed Air
Transducer ETV Computer
Dryer
High-pressure Air
Tank

Physical Connection DSP-based VAM

Thick Pipe Path
Controller
Thin Pipe Path Electronic Signals

Fig. 2. System configuration of the VAM pneumatic servo system.

thermal dynamic analysis, the instantaneous temperature Tin and pressure Pin of the intake chamber of the VAM can be
expressed as follows [4,19]:
Ein
T in 3
min  C V

min RT in REin
Pin ; 4
V in V in C V
where subscript in represents the intake chamber of VAM, and C V represents the specific heat ratio at a constant volume. The
trapped mass min , inner energy Ein , and instantaneous chamber volume V in can be derived using the mass, energy, and vol-
ume conservation laws as follows:
Z
min t min;init _ in tdt
m 5

Z
Ein t Ein;init E_ in tdt 6

Z
V in t V_ in tdt; 7

_
where min;init and Ein;init are the initial mass and energy in the intake chamber of the VAM. The volume change rate Vt varies
according to the mechanical angle of the rotor and the initial volume of the intake chamber is assumed as zero. The rate of
the energy flow through the valve can be represented as follows:

E_ in C p T in m
_ in ; 8

where C p represents the specific heat at a constant pressure.

Similarly, the instantaneous temperature Tex and pressure Pex of the exhaust chamber of the VAM can be expressed as
follows:
Eex
T ex 9
mex  C V

mex RT ex REex
Pex ; 10
V ex V ex C V
where subscript ex represents the exhaust chamber of the VAM. The parameters mex , Eex , and V ex can be obtained as follows:
116 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128

Z
mex t mex;init  _ ex tdt
m 11

Z
Eex t Eex;init  E_ ex tdt 12

Z
V ex t V_ ex tdt; 13

where mex;init and Eex;init are the initial mass and energy in the exhaust chamber of the VAM. Thus, the instantaneous energy in
the intake chamber can be expressed using Eqs. (1), (2), (6), and (8) as follows:
Z sq
C p C ex T in P in / 2c c1
Unchoked : Ein t Ein;init p Pex =Pin 2=c  Pex =Pin c dt 14
T in Rc  1

Z  c1=2c1
C p C ex T in Pin / 1=2 2c
Choked : Ein t Ein;init p c dt 15
RT in c1
For the sake of simplicity, the instantaneous energy of both the unchoked and choked conditions can be represented as
follows:
Z
Ein t Ein;init Uin P in ; Pex ; T in ; t /dt; 16

where Uin is a complicated nonlinear function of the time-varying parameters Pin , Pex , and T in .
The pneumatic energy is ultimately transferred to the mechanical force that is applied against the rotor inertia and exter-
nal load. Therefore, the difference in the air pressures in the intake and exhaust chambers provides the output torque, sam
required to move the rotor [16]

sam t Pin t; /V in t; h  Pex tV ex t; h

17
J am ht bam ht
_ sld t;

where h, J am , bam , and sld are the mechanical angle of the rotor, equivalent rotational inertia, damping coefficient, and external
load torque of the VAM, respectively. Notably, the pressure of the intake chamber P in t; / is a function of time and control-
lable EOA of the ETV. However, the exhaust chamber pressure Pex t is a function of time only because the EOA of the exhaust
port is a constant. As a result, the second-order VAM pneumatic servo system for the rotor angle h can be rewritten as follows
using Eqs. (4), (10), and (17):
 
h J 1 R
am Ein  Eex   bam h_  sld 18
CV
Substituting (16) into (18) yields the following expression:
 Z 
h J 1 R
am /dt  Eex   bam h_  sld
Ein;init Uin
CV
Z
J 1 _ J 1
U /dt J am C V 1 REin;init  Eex   J 1
b
am am h C
am V R in am sld
Z 19
Ah_ B /dt d

Ah_ Bu d;
1
where A J 1
am bam < 0 and B J am C V RUin > 0 are the system parameters, d is a bounded time-varying uncertainty
R
defined as d J C REin;init  Eex  sld , h_ x is the rotor speed, and u is the control signal defined as u /dt (i.e., u_ /).
1 1
am V
In Eq. (19), the control signal of the system is u, which is the integral of the EOA / of the ETV (i.e., u_ /). As only the EOA /
is controllable in a practical system, the objective of the control system design is to derive the differential of the control sig-
nal u_ so that the actual control input signal / can be obtained, thereby rendering the direct control of the EOA feasible. Thus,
the input MAF of the compressed air can be regulated, and the corresponding pressure and torque can be controlled to
change the speed of the VAM.

3. Problem formulation

According to Eq. (19), the VAM is a highly nonlinear and time-varying system. One of the difficulties in controlling VAM
systems is the presence of uncertainties or vagueness in their parameters. Moreover, most of the system parameters, such as
 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128 117

Ein;init , Eex , and Uin , are very difficult to measure during practical implementation. Assuming a nominal model of the VAM
pneumatic servo system, Eq. (19) can be represented as follows:

x_ An x Bn u d; 20

where An and Bn are the nominal values of A and B, respectively. If parameter variations and uncertainty are involved, the
system model in Eq. (20) can be formulated as

x_ An DAx Bn DBu d
21
An x Bn u L;

where DA and DB denote the time-varying parameter variations and L is the lumped uncertainty, which is defined as
L DAx DBu d.
At this point, the DSMC system is designed to achieve the speed tracking control of the VAM pneumatic servo system.
First, the speed tracking error e xd  x is defined, where xd represents the speed command. Subsequently, a hierarchical
sliding surfaces structure, which combines the first double integral sliding surface s and the secondary dynamic sliding sur-
face q [23], is designed as follows:
Z t Z tZ t
s e c1 esds c2 esdsds 22
0 0 0

q s_ c3 s; 23

where c1 , c2 , and c3 are all positive constants and chosen to correspond to the coefficients of the Hurwitz polynomials. Dif-
ferentiating q with respect to time yields the following expression:

q_ t s c3 s_
 Z t 
x _  Bn u_  L_ c1 e_ c2 e c3 x
d  An x _ d  An x  Bn u  L c1 e c2 esds
0 24
Z t
d c3 x
Xx _ d  Bn u_ c1 e_ p1 e p2 _
esds  c3 L L;
0

_  c3 An x  c3 Bn u. There-
where p1 , p2 , and X are defined, for the sake of simplicity, as p1  c1 c3 c2 , p2  c2 c3 , and X  An x
fore, the asymptotic stability of the DSMC system is ensured if the control signal u is applied to the control law uDSMC of the
DSMC system, expressed as Eqs. (25) and (26):
Z t
uDSMC u_ DSMC sds 25
0

 Z t 
u_ DSMC B1
n X
x d c 3 _
x d c 1 _
e p1 e p2 es ds dsgn q ; 26
0

where d is the hitting control gain. Assuming the positive upper bounds of L and L_ are b1 and b2, respectively, the uncertainty
of the DSMC system shown in Eq. (24) satisfies the following:

_ 6 c3 kLk kLk
kc3 L Lk _ 6 c3 b1 b2 : 27
If the hitting control gain is chosen as follows:

d > c3 b1 b2 28
Therefore, it is big enough to force the system state to approach and remain at the secondary dynamic sliding surface q. If
q 0, then s_ c3 s 0. Therefore, the first double integral sliding surface is s 0. Moreover, if
Rt Rt Rt
s e c1 0 esds c2 0 0 esdsds 0, the speed tracking error is e 0 when the parameters c1 and c2 are appropriately
chosen. The control diagram of the VAM pneumatic servo system using the DSMC system is shown in Fig. 3.
_ can be derived by controlling the EOA / of
According to the design of the hierarchical sliding surfaces, the control law, u,
the ETV directly, as shown in Eq. (26). In addition, the control law, u, can asymptotically stabilize the VAM pneumatic servo
system, as shown in Eq. (25). However, practical implementation of the control law (Eq. (26)), as shown in Fig. 3, is extremely
difficult because the system parameter X is always unknown during the control process. Moreover, because the upper
bounds of b1 and b2 are always unclear, the hitting control gain d should always be adjusted through trial-and-error proce-
dures when the robustness of the control system decreases. In addition, the constant control gains of the DSMC system can-
not ensure stable performance of the control because of the highly nonlinear and time-varying control characteristics of the
VAM.
118 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128

DSMC System

First Second
d
DSMC
Sliding dt Sliding Control u DSMC
d + Surface s Surface of ETV
e Law
s
(26)
(22) (23)
VAM

DSP-based VAM controller

Fig. 3. Control diagram of the VAM pneumatic servo system using the DSMC system.

4. Proposed observation-based ADSMC system

For the proposed observation-based ADSMC system, a PIDFNN observer is used to imitate an ideal DSMC as the main
tracking controller, and the supervisor compensator is designed to eliminate the approximation error between the ideal
DSMC and the PIDFNN observer. Because the main control effort is directly obtained through the PIDFNN observer, the sys-
tem parameters, bound information, and hitting control are unnecessary in the ADSMC system. Moreover, the adaptive algo-
rithms for tuning the control parameters are derived through the Lyapunov theory to ensure the stability of the system. Thus,
the robustness of the DSMC system and the online learning ability of the PIDFNN can be integrated into an ADSMC system to
accurately control the VAM pneumatic servo system.

4.1. PIDFNN observer

The PIDFNN, as shown in Fig. 4, is used to imitate the DSCM system. The PIDFNN has a five-layer architecture and an aux-
1
iliary PIDNN structure [16]. In layer one, the ith node output yi is ei , where e1 and e2 are the speed tracking error, e, and its
_ respectively. In layer two, each node uses a Gaussian function to act as a fuzzy membership function:
derivative, e,
8 h i2 9
>
< yi1  mik > =
2
yik exp  ; k 1; 2;    ; n; 29
>
: rik
2 >
;

y ( 5)
Layer 5
wk
y1( 4) y2( 4) yn( 4 )
Layer 4 1

1 y1 y2 yn
( 3) ( 3) ( 3)
Layer 3 y 1 y 2 y n
...

jk
1
o1 1 o2 o3
( 2) ( 2) ( 2)
Z 1
1
( 2) ( 2) ( 2)
Layer 2 y11 y12 y1n y21 y22 y2n Z 1

... ...
1
(1) 1 1
y
1 y2(1)
Layer 1

PIDNN
e1 e2
Fig. 4. Structure of the adopted PIDFNN observer.
 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128 119

2
where yik is the output of layer two, and mik and rik are the mean and standard deviation of the Gaussian functions, respec-
tively. In layer three, the node means the preconditioning part of the fuzzy rules. Consequently, the node in layer three is
Q
denoted by , which multiplies the incoming data from layer two. The output of the kth node in layer three is:

3
Y
2
2
yk yik ; 30
i1

3
where yk is the firing strength of the kth fuzzy rule. In layer four, the node performs a product operation by multiplying the
^k , and y3
incoming data, y k :

4 3
X
3
yk yk y ^k y3
k fjk oj ; 31
j1

4
^k is the output of the
where yk is the output of layer four. f1k , f2k , and f3k are the P, I, and D control gains of the PIDNN, y
auxiliary PIDNN using the control gain fik and oj performs the P, I, and D calculations for the speed tracking error as follows:
o1 N e1 N e2 N 32

o2 N e1 N e2 N o2 N  1 33

o3 N e1 N e2 N  e1 N  1  e2 N  1; 34
4
where N represents the Nth iteration. In layer five, the node is a linear combination of the outputs yk from layer four, con-
sidering different weights, wk , which act as a defuzzifier:
X
n
4
X
n
3
y5 wk yk wk yk y ^k ; 35
k1 k1

where y5 is the final output of the PIDFNN and wk is the connective weight.
To complete the controller design, the vectors W, Y, m, r, and f are defined as the collection vectors of parameters:

W w1 ; w2 ; . . . ; wk ; . . . ; wn T 2 Rn1 36

4 4 4 T
Y y1 ; y2 ; . . . ; yk ; . . . ; yn4  2 Rn1 37

m m11 ; m12 ; . . . ; mik ; . . . ; m2n T 2 R2n1 38

r r11 ; r12 ; . . . ; rik ; . . . ; r2n T 2 R2n1 39

f f11 ; f12 ; . . . ; fjk ; . . . ; f3n T 2 R3n1 40

Therefore, the output of the PIDFNN observer can be represented in vector form:

uPIDFNN y5 W T Ym; r; f 41

4.2. ADSMC system

By assuming that the lumped uncertainty and its differential are known, the ideal DSMC system, according to Eqs. (25)
and (26), can be defined as follows:
Z t
uDSMC u_ DSMC sds 42
0

 Z t 
u_ DSMC B1 X
x d c _
x
3 d c 1 _
e p e p es ds  c 3 L _
L 43
n 1 2
0

By using the approximation theory, there exists an ideal PIDFNN, uPIDFNN , such that it can uniformly approximate the ideal
DSMC system, u_ DSMC , closely:

u_ DSMC uPIDFNN e W T Ym ; r ; f e; 44

where e is a minimum reconstructed error caused by the insufficient structure of the PIDFNN observer and W , m , r , and f   

are the optimal parameter vectors of W, m, r, and f, respectively. However, the optimal uPIDFNN cannot be obtained; therefore,
120 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128

an online estimating PIDFNN observer u ^ PIDFNN is used to approximate the optimal PIDFNN observer. Moreover, because of the
imperfect approximation, a supervisor compensator ur is proposed to attenuate the approximation error between the ideal
DSMC system, u_ DSMC , and the practical PIDFNN observer, u ^ PIDFNN . Thus, the proposed ADSMC system is defined as:
Z t
uADSMC u_ ADSMC sds 45
0

^ T Ym;
^ PIDFNN ur W
u_ ADSMC u ^ r^ ; ^f ur ; 46

where u_ ADSMC is the control variable, i.e., the EOA / of the ETV connected to the intake pipe, and W,^ m,^ r^ , and ^f are the esti-
mations of the optimal parameter vectors W  , m , r , and f , respectively. For ease of notation, define Y   Ym ; r ; f and
^  Ym;
Y ^ r ^ ; ^f in the following derivation.
The estimation error, u e , of the control signal can be denoted as follows:

^ PIDFNN W T Y   W
e u_ DSMC  u
u ^ e
^ TY
47
W e W
fT Y ^ TY
eW
fTY
^ e;

where W f W  W ^ and Y
e Y   Y.
^ Moreover, to increase the speed of the parameter convergence, the optimal weight
vector, W  , is decomposed into two parts [32]:
W  gW P W P gW I W I ; 48
Rt
where gW P and gW I are both positive constants, W P and W I are the P and I parts of W  , respectively, and W I 0
W P ds. Fur-
thermore, the estimated weight vector, W, ^ is also decomposed into two parts:
^ g W
W ^ P g W ^ I; 49
WP WI
Rt
^ P and W
where W ^ I are the P and I parts of W,
^ respectively, and W
^I W f can be expressed as
^ P ds. Therefore, W
0

f g W
W f I g W  g W ^ P; 50
WI WP P WP

f I W  W
where W ^ I . Therefore, the expansion of Y
e in the Taylor series can be obtained as follows:
I
2 3 2 3 2 4
3
2 4
3 4
@y1
4
@y1 @y1
e
y1 6 @mT 7 6 @ rT 7 6 @f 7 T

6 7 6 @y4 7 6 @y4 7 6 4 7
6 e7 6 2 7
4 6 2 7 6 @y2 7
6 y2
7 6 7 6 7 6 @fT 7
e 6 ^ 6
^ 6 @ r 7jrr^ r  r 7j ^ f  ^f Q  Y T f Te
mm Yrr Yf f Q
T e
Y 7 6 @m 7jmm^ m  m
T T
6 7 ff 51
6 .. 7 6 . 7 6 . 7 6 .. 7
4 . 5 6 .. 7 6 .. 7 6 . 7
4
4 5 4 5 4 4 5
e
yn 4
@yn
4
@yn @yn
@mT @ rT @fT

4
h4 4
i h 4 4 4
i h 4 4 4
i
where Y m @y1 @y2    @yn jmm^ 2 R2nn , Y r @y1 @y2    @yn jr^r 2 R2nn , Y f @y1 @y2    @yn jf^f 2 R3nn ,
@m @m @m @r @r @r @f @f @f
f ^ r
m m  m, e r  r^ , and e
f f  ^f. Q 2 Rn1 is a higher-order term vector. Substituting (51) into (47) yields
e W
fT Y ^ T Y T f Te fT ^
mm Yrr Yf f Q W Y e
eW
u T e
52
e W
fT Y ^ TYT f ^T Te ^ T Te ^T fT ^
W mm W Yrr W Yf f W Q W Y e

Applying Eq. (50) gives the following expressions:

e W
fT Y ^ TYT f ^T Te ^ T Te ^T f ^ T^
m m W Y r r W Y f f W Q gW I W I gW P W P  gW P W P Y e

eW
u
^ TYT f ^T Te ^ T Te fT ^ ^T^
m m W Y r r W Y f f gW I W I Y  gW P W P Y h
W 53
fT ^ r
m Y mW ^ ef T Y f W
e T Y rW ^ g W fTY^ g W ^ TY^ h;
WI I WP P

where the approximation error term h W e W

fT Y ^ T Q g W T Y
^ e. This allows Theorem 1, below, to be obtained.
WP P

Theorem 1. Consider the VAM pneumatic servo system represented using Eq. (21). If the proposed ADSMC system represented
using Eqs. (45) and (46) comprises a PIDFNN observer u ^ PIDFNN , represented using Eq. (41) with adaptive laws (54)(58), and a
supervisor compensator ur , which is represented using Eq. (59) with compensation law (60), is adopted as the control signal u_ ADSMC
for controlling the EOA / of the ETV, then the asymptotic stability of the VAM pneumatic servo system can be ensured:
^ P qY
W ^ 54

^_ I qY
W ^ 55
 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128 121

^_ gm qY m W
m ^ 56

r^_ gr qY r W
^ 57

^f_ g qY f W
^ 58
f

^
ur h 59

^_ g q;
h 60
h

^ is an online estimated value of h.

where gm , gr , gf , and gh are the positive learning rates, and h

Proof. The Lyapunov function for the ADSMC system is selected as follows:
1 2 gW I f T f 1 1 1 eT e 1 e2
V q WI WI f
mT f
m r
eT r
e f f h ; 61
2Bn 2 2 gm 2 gr 2 gf 2 gh
e h  h.
where the approximation error is defined as h ^ By differentiating Eq. (61) and using Eqs. (46), (47), and (53), one can
get the following:

^_ I  1 f 1 T_ 1 e ^_
^  ef T ^f_  h
1
V_ qu_ DSMC  u_ ADSMC  gW I W
fTW
I
^_ 
mT m r
e r h
gm gr gf gh
q f ^ r
mT Y mW ^ ef T Y f W
e T Y rW ^ g W fTY^ g W ^ TY ^ g W
^ h  h fTW^_ I
WI I WP P WI I

1 1 T_ 1 e ^_
^  ef T ^f_  h
1
 f ^_ 
mT m r
e r h
gm gr gf gh
! !
^_ ^_
f T qY
gW I W ^_ I  g qW
^ W ^ TY^ f ^ m
m T qY m W r
eT ^  r
qY r W
I WP P
gm gr
! 0 1
^f_ ^_
h
^ 
ef T qY f W e @q 
h A 62
gf gh

Therefore, if the adaptive laws of the PIDFNN observer are expressed as Eqs. (54)(58) and the compensation law of the
supervisor compensator is expressed as Eq. (60), then Eq. (62) can be rewritten as:
_
Vt ^TY
gW P q2 Y ^ 60 63

Because V_ 6 0, V_ is negative semi-definite, i.e., Vt 6 V0, implying that qt, W

f I, f f, and e
e, e
m, r h are bounded. By defin-
^ T^ _
ing the function Kt  g q tY Y 6 Vt and integrating the function Kt with respect to time, one can get the
2
WP
following:
Z t
Ks ds 6 V0  Vt 64
0

Because V0 is bounded and Vt is non-increasing and bounded, the following result can be obtained:
Z t
lim Ks ds < 1 65
t!1 0

Because K _ t is bounded, Kt is uniformly continuous. When Barbalats lemma [33] is applied, limt!1 Kt 0. This
implies that the secondary sliding surface, q, converges to zero as t ! 1. If q 0, s_ c3 s 0. When c3 is a positive constant,
Rt Rt Rt
limt!1 s 0. Similarly, if the first sliding surface s 0 and e c1 0 esds c2 0 0 esdsds 0, limt!1 e 0 when c1 and c2
match the Hurwitz condition. Therefore, the asymptotical stability of the proposed ADSMC system is ensured. Fig. 5 illus-
trates the configuration of the proposed ADSMC.
Regarding the online learning algorithms of the PIDFNN derived through the gradient descent method [16], the adaptive
laws of the ADSMC system derived through the Lyapunov stability theory further use the dynamic sliding surface variable, q,
to perform a faster tracking error convergence, as shown in Eqs. (54)(58). This approach is effective in changing the control
parameters of the ADSMC system to realize more precise and stable tracking performance levels. h
122 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128

ADSMC System
Supervisor ur + ADSMC
Compensator Control Law
(59) + (46)
Compensation

h
First d
Second u ADSMC
Sliding
dt
Sliding
Law
d + e Surface s Surface (60)
of ETV
s
Adaptive
(22) (23)
Laws
VAM
(54)-(58)

PIDFNN
d Observer
dt u PIDFNN
(29)-(41)
  
, m ,  ,
WP , W I

DSP-based VAM controller

Fig. 5. Control diagram of the VAM pneumatic servo system using the proposed ADSMC system.

Piston Actuated Electropneumatic Digital

Valve Transducer Oscilloscope

Pressure
Regulator

Compressor Personal Filter

Computer DSP-based
High-pressure VAM
Air Tank Controller

Compressed Air VAM Encoder

Dryer
Magnetic
Powder Brake

Fig. 6. Experimental setup of the VAM pneumatic servo system.

5. Experimental result

5.1. Experimental setup

Fig. 6 shows the practical experimental setup of the VAM pneumatic servo system. To perform high performance real-
time control for the VAM pneumatic servo system, a resultant big control program loop processing data streams should
be executed in a small time interval. Since the control algorithm involves some complicated arithmetical operations as
shown in Fig. 5, a DSP rather than a microcontroller is chosen to implement the developed control algorithm in this study.
In this regard, the developed DSP control chip with online observation and learning abilities is suitable for controlling the
plants without clear system parameters such as damping and spring coefficients as well as uncertainty information such
as friction and external disturbance. Moreover, a backward difference approximation method and a trapezoidal integration
method are used for the digital realization of differential and integral calculus, respectively, in the DSP [34].
The DSP-based VAM controller has a 150 MHz CPU frequency, 14-bit resolution analog-to-digital converter, 14-bit reso-
lution digital-to-analog converter (DACs), and an encoder interface (quadrature encoder pulse, QEP). In addition, the pres-
sure regulator regulates the pressure of the MAF to 7 kgf/cm2. The proposed control systems were realized on personal
 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128 123

Main QEP ISR

Parameters Encoder Rotor Angle

Initialization Interface Calculation

I/O Pulses Rotor Speed

Initialization Accumulation Calculation

Interrupt END Speed Tracking

Service Error and Its
Routine 0.025ms Differentiation
(ISR) Setting Calculations
e, e
Enable Execution of
Interrupt Controllers

Monitor Data Online Tuning

of Controllers

Disable
D/A Conversion
Interrupt

END
END
1ms

Fig. 7. Control strategy flowchart.

computer by using C language and then downloaded to the flash ROM of the DSP. The methodology proposed for real-time
control implementation comprises one main program and one interrupt service routine (ISR) as shown in Fig. 7. In the main
program, parameters and I/O initializations are set first and then the interrupt interval for the ISR is set. After the interrupt is
enabled, the ISR, with a 1 ms execution frequency, first calculates the rotor angle and speed from the encoder interface
(quadrature encoder pulse, QEP) and then calculates the speed tracking error and its differentiation. Subsequently, the con-
trol signals for the EOA / of the ETV are determined using the control systems. Finally, the control signals are sent to the ETV
through a DAC to control the MAF rate of the delivered compressed air for the speed tracking.

5.2. Control performance measures and comparison

To investigate the speed tracking performance and robustness of the proposed ADSMC system, the VAM pneumatic servo
system was tested under a nominal case (Case 1) and a payload case (Case 2). In Case 2, a magnetic powder brake was set at a
5 Nm torque to simulate the external payload. The maximum speed tracking error, average speed tracking error, and stan-
dard deviation of the speed tracking error were also evaluated for comparison with the values obtained from earlier studies
[16]; in total, five controllers, including PID, PIDNN [21], FNN [22], and PIDFNN [16], were tested and subsequently
compared.
To compare the control performance fairly, the inputs for the PIDNN, FNN, PIDFNN controllers and the PIDFNN observer
within the ADSMC system were all set to speed tracking error, e, and its derivative, e._ Moreover, the number of fuzzy rule
nodes in the FNN was set to nine. Regarding the additional subnetwork computation, seven fuzzy rule nodes were designed
in the PIDFNN and PIDFNN within the ADSMC system so that they could be compared. In addition, all the constant control
parameters, such as the PID gains, sliding surface parameters, and learning rates, were obtained by trial-and-error methods
to achieve the best transient control performance levels regarding the possible occurrences of the uncertainties and require-
ment of steady-state stabilities. On the other hand, all the adaptive control parameters, such as the fuzzy membership func-
tion parameters and the connective weights, were obtained using the initial formation strategy. In the initial formation, the
membership function parameters were selected uniformly and the other adaptive control parameters were selected ran-
domly. Subsequently, all the adaptive control parameters were saved after one cycle of speed tracking for the initial forma-
tion. All the control systems using well-initialized parameters could then be fairly compared for their optimal set points in
the experiments.

5.3. Experimentation

In the experiments, the control gains of the PID controller were set to Kp = 15, Ki = 2, and Kd = 0.03. For the fuzzy
membership functions of the ADSMC system, the initial means were 15, 10, 5, 0, 5, 10, and 15, and the initial standard
124 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128

1200
100rpm Speed
1100

Rotor speed (rpm)

Command 10s
1000
900
800
700
600
500 Rotor Speed
400
Time (s)
a
100
Tracking error (rpm)

75
10s
50
25
0

25rpm
Time (s)
b
4
3.5
Control signal (V)

3
2.5
2
1.5
1
10s
0.5
0.5V
0
Time (s)
c
4 w1 m14 h
Adaptive parameters

14 11

11
3

2
w1 m14
1
14
0

-1
h
10s
Time (s)
d
Fig. 8. Experimental results of the VAM pneumatic servo system using the ADSMC system for Case 1: (a) speed tracking response; (b) speed tracking error;
(c) control signal; (d) adaptive control parameters.

deviations were all 2. The constant control parameters for the ADSMC system were selected as c1 0:075, c2 0:03, c3 0:2,
gW P 0:4, gW I 0:5, gm 0:1, gr 0:2, gf 0:2, gh 0:05, and n 7. For practical implementation, because the weight
vector W significantly dominates the network output, their learning rates, i.e., gW P and gW I , are always larger than those
of the others to achieve agile observations.
Figs. 8 and 9 show the experimental results of the VAM pneumatic servo system using the ADSMC system for Cases 1 and
2, respectively. According to the experimental results, the EOA of the ETV was adjusted to change the MAF for tacking the
speed command accurately. Furthermore, greater MAFs were required to manage the external load torque in Case 2. Accord-
ing to the adaptive laws developed in Eqs. (54)(58) and the compensation law in Eq. (60), the network parameters can be
varied to confront the uncertainties during the control process, as shown in Figs. 8(d) and 9(d).
 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128 125

1200
100rpm Speed

Rotor speed (rpm)

1100
Command 10s
1000
900
800
700
600
500 Rotor Speed
400
Time (s)
a
100
Tracking error (rpm)

75
10s
50
25
0
-25
-50

25rpm
Time (s)
b
4
Control signal (V)

3.5
3
2.5
2
1.5
1
10s
0.5
0 0.5V
Time (s)
c
4 w1 m14 h
14 11
Adaptive parameters

11
3
w1
2
m14
1
14
0

-1 h
10s
Time (s)
d
Fig. 9. Experimental results of the VAM pneumatic servo system using the ADSMC system for Case 2: (a) speed tracking response; (b) speed tracking error;
(c) control signal; (d) adaptive control parameters.

Figs. 10 and 11 show comparisons between the speed tracking responses and errors of the VAM pneumatic servo system
using the PID, PIDNN, FNN, PIDFNN, and ADSMC controllers in Cases 1 and 2, respectively. As indicated by the experimental
results, the tracking performances in Case 2, as shown in Fig. 11, are not as high as those in Case 1, as shown in Fig. 10; this is
because of the heavy inertias resulting from the additional payload disturbance. For the PID controller, the tracking errors in
Case 2 are much greater than those in Case 1 because the constant control gains cannot address the additional payload.
Though the unsatisfactory tracking errors can be reduced with a larger control gains for Case 2, the steady-state errors will
be consequently exacerbated in Case 1. Furthermore, using the PIDNN, FNN, PIDFNN, and ADSMC controllers can yield dif-
ferent levels of improvement owing to their dissimilar adaptive control abilities. Compared with the other controllers, the
adaptive laws of the proposed ADSMC further involve the dynamic sliding surface variable q, which markedly improves
126 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128

1100
Speed Command PID PIDNN FNN PIDFNN ADSMC

Rotor speed (rpm)

1000

900
PIDFNN PID
800 PIDNN
ADSMC FNN
700

600

500
0 10 20 30 40 50 60 70 80 90 100
Time (s)
a

60 PID PIDNN FNN PIDFNN ADSMC

Tracking error (rpm)

PID PIDNN
40
FNN
20

0
ADSMC
-20
PIDFNN
-40

-60
0 10 20 30 40 50 60 70 80 90 100
Time (s)
b
Fig. 10. Comparison of VAM pneumatic servo system using various control systems: (a) speed tracking responses in Case 1, (b) speed tracking errors in Case
1.

1100
Speed Command PID PIDNN FNN PIDFNN ADSMC
Rotor speed (rpm)

1000

900
PIDFNN PID
800 ADSMC PIDNN
FNN
700

600

500
0 10 20 30 40 50 60 70 80 90 100
Time (s)
a

60 PID PIDNN FNN PIDFNN ADSMC

PID
Tracking error (rpm)

PIDNN
40
FNN
20

0
ADSMC
-20
PIDFNN
-40

-60
0 10 20 30 40 50 60 70 80 90 100
Time (s)
b
Fig. 11. Comparison of VAM pneumatic servo system using various control systems: (a) speed tracking responses in Case 2, (b) speed tracking errors in Case
2.
 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128 127

Table 1
Performance measures of VAM pneumatic servo system using various control systems in Case 1.

Tracking errors (rpm) Controllers

PID PIDNN FNN PIDFNN ADSMC
Maximum 45 41 34 32 27
Average 26.65 23.69 18.14 10.44 9.52
Standard deviation 9.868 8.607 7.17 5.93 5.2

Table 2
Performance measures of VAM pneumatic servo system using various control systems in Case 2.

Tracking errors (rpm) Controllers

PID PIDNN FNN PIDFNN ADSMC
Maximum 54 48 38 36 31
Average 26.87 22.13 18.49 12.10 11.29
Standard deviation 12.95 11.66 7.6 6.95 6.04

the tracking error convergence and reduces the steady-state errors. Although selecting higher learning rates for the PIDNN,
FNN, and PIDFNN controllers can achieve shorter transient response time; learning rates that are too high may cause the
rotor to generate unstable oscillations in the steady-state. As seen in Figs. 10 and 11, the experiments reveal that the pro-
posed ADSMC system delivers the greatest tracking performance.
To further evaluate the control performance of the VAM pneumatic servo system using ADSMC system, the speed-
tracking performance measures of the PID, PIDNN, FNN, PIDFNN, and ADSMC controllers in Cases 1 and 2 were compared
in Tables 1 and 2, respectively. It is clear that the speed tracking was much better in the ADSMC system. The ADSMC system
exhibited superior performance when compared with the other four controllers in both test cases.

6. Conclusion

This study demonstrated the development and application of an observation-based ADSMC system for controlling the
rotor speed of a highly nonlinear and time-varying VAM pneumatic servo system by directly regulating the EOA of the
ETV. The configuration and operating principles of the VAM pneumatic servo systems were introduced as a first step to
derive the dynamic model of the VAM. Subsequently, the theoretical bases of the proposed ADSMC system using a PIDFNN
observer and a supervisor compensator with learning algorithms were described in detail. The proposed ADSMC was found
to combine the robustness of the DSMC and the online learning ability of the PIDFNN. Experiments were then conducted
using DSP to test the effectiveness of the proposed ADSMC system. Experimental results clearly illustrate the improved
speed-tracking performance of the proposed ADSMC system for a VAM pneumatic servo system. In summary, the main
advantages of the proposed ADSMC system are: (a) easy implementation without system parameters and uncertain bound-
ary information; (b) high robustness against uncertainties; and (c) high precision and stable tracking performance levels
without chattering phenomena.

Acknowledgements

The authors would like to acknowledge the financial supports of the National Taiwan Normal University (NTNU) and the
Ministry of Science and Technology in Taiwan, R.O.C. under Grant no. MOST 105-2221-E-003-014.

References

[1] R. Saidur, N.A. Rahim, M. Hasanuzzaman, A review on compressed-air energy use and energy savings, Renew. Sustain. Energy Rev. 14 (2010) 1135
1153.
[2] S.R. Padian, F. Takemura, Y. Hayakawa, Control performance of an air motor-can air motors replace electric motors?, in: Proc. Int. Conf. Robotics &
Automation, Detroit, MI, May 1999, pp. 518524.
[3] M. Sorli, L. Gastaldi, E. Codina, S. de las Heras, Dynamic analysis of pneumatic actuators, Simul. Pract. Theory 7 (1999) 589602.
[4] M. Karpenko, N. Sepehri, Development and experimental evaluation of a fixed-gain nonlinear control for a low-cost pneumatic actuator, Proc. IEE
Control Theory Appl. 153 (2006) 629640.
[5] Y.C. Tsai, A.C. Huang, FAT-based adaptive control for pneumatic servo systems with mismatched uncertainties, Mech. Syst. Signal Process. 22 (2008)
12631273.
[6] S. Riachy, M. Ghanes, A nonlinear controller for pneumatic servo systems: design and experimental tests, IEEE/ASME Trans. Mechatron. 19 (2014)
13631373.
[7] G. Niu, Y. Zhao, M. Defoort, M. Pecht, Fault diagnosis of locomotive electro-pneumatic brake through uncertain bond graph modeling and robust online
monitoring, Mech. Syst. Signal Process. 5051 (2015) 676691.
[8] Z.L. Zhao, Z.C. Qiu, X.M. Zhang, J.D. Han, Vibration control of a pneumatic driven piezoelectric flexible manipulator using self-organizing map based
multiple models, Mech. Syst. Signal Process. 7071 (2016) 345372.
128 S.-Y. Chen, S.-S. Gong / Mechanical Systems and Signal Processing 94 (2017) 111128

[9] J.M. Bajkowski, B. Dyniewicz, C.I. Bajer, Semi-active damping strategy for beams system with pneumatically controlled granular structure, Mech. Syst.
Signal Process. 7071 (2016) 387396.
[10] K. Xing, J. Huang, Y. Wang, J. Wu, Q. Xu, J. He, Tracking control of pneumatic artificial muscle actuators based on sliding mode and non-linear
disturbance observer, IET Control Theory Appl. 4 (2010) 20582070.
[11] Y.T. Shen, Y.R. Hwang, Design and implementation of an air-powered motorcycles, Appl. Energy 86 (2009) 11051110.
[12] J. Naranjo, E. Kussul, G. Ascanio, A new pneumatic vanes motor, Mechatronics 20 (2010) 424427.
[13] C.H. Lu, Y.R. Hwang, Y.T. Shen, Backstepping sliding mode tracking control of a vane-type air motor X-Y table motion system, ISA Trans. 50 (2011) 278
286.
[14] Y.R. Hwang, C.H. Lu, Modeling of an air motor servo system and robust sliding mode controller design, J. Mech. Sci. Technol. 26 (2012) 11611169.
[15] Y.R. Hwang, S.Y. Huang, System identification and integration design of an air/electric motor, Energies 6 (2013) 921933.
[16] S.Y. Chen, Y.H. Hung, S.S. Gong, Speed control of vane-type air motor servo system using proportional-integral-derivative-based fuzzy neural network,
Int. J. Fuzzy Syst. (2016), http://dx.doi.org/10.1007/s40815-015-0134-0.
[17] J. Wang, X. Luo, L. Yang, L.M. Shpanin, J. Nan, S. Mangan, J.W. Derby, Mathematical modeling study of scroll air motors and energy efficiency analysis-
part I, IEEE/ASME Trans. Mech. 16 (2011) 112121.
[18] J. Wang, X. Luo, L. Yang, L.M. Shpanin, J. Nan, S. Mangan, J.W. Derby, Mathematical modeling study of scroll air motors and energy efficiency analysis-
part II, IEEE/ASME Trans. Mech. 16 (2011) 122132.
[19] Y.H. Hung, Y.M. Tung, H.W. Li, A real-time model of an automotive air propulsion system, Appl. Energy 129 (2014) 287298.
[20] M.H. Khooban, T. Niknam, M. Sha-Sadeghi, Speed control of electrical vehicles: a time-varying proportionalintegral controller-based type-2 fuzzy
logic, IET Sci. Meas. Technol. 10 (2016) 185192.
[21] S.Y. Chen, F.J. Lin, Decentralized PID neural network control for five degree-of-freedom active magnetic bearing, Eng. Appl. Artif. Intell. 26 (2013) 962
973.
[22] F.J. Lin, P.K. Huang, W.D. Chou, Recurrent fuzzy neural network controlled linear induction motor servo drive using genetic algorithms, IEEE Trans. Ind.
Electron. 54 (2007) 14491461.
[23] A.J. Koshkouei, K.J. Burnham, A.S.I. Zinober, Dynamic sliding mode control design, IEE Proc. Control Theory Appl. 152 (2005) 392396.
[24] M.R. Soltanpour, P. Otadolajam, M.H. Khooban, Robust control strategy for electrically driven robot manipulators: adaptive fuzzy sliding mode, IET Sci.
Meas. Technol. 9 (2015) 322334.
[25] M.R. Soltanpour, M.H. Khooban, M.R. Khalghani, An optimal and intelligent control strategy for a class of nonlinear systems: adaptive fuzzy sliding
mode, J. Vib. Control 22 (2016) 159175.
[26] M.H. Khooban, T. Niknam, F. Blaabjerg, Free chattering hybrid sliding mode control for a class of non-linear systems: electric vehicles as a case study,
IET Sci. Meas. Technol. 10 (2016) 776785.
[27] M. Rahmani, A. Ghanbari, M.M. Ettefagh, Robust adaptive control of a bio-inspired robot manipulator using bat algorithm, Expert Syst. Appl. 56 (2016)
164176.
[28] M. Rahmani, A. Ghanbari, M.M. Ettefagh, A novel adaptive neural network integral sliding-mode control of a biped robot using bat algorithm, J. Vib.
Control (2016), http://dx.doi.org/10.1177/1077546316676734.
[29] B. Li, H. Du, W. Li, Fault-tolerant control of electric vehicles with in-wheel motors using actuator-grouping sliding mode controllers, Mech. Syst. Signal
Process. 7273 (2016) 462485.
[30] M. Rahmani, A. Ghanbari, M.M. Ettefagh, Hybrid neural network fraction integral terminal sliding mode control of an Inchworm robot manipulator,
Mech. Syst. Signal Process. 80 (2016) 117136.
[31] VAM Data sheet, GAST, <http://www.gastmfg.com/>.
[32] C.F. Hsu, Y.C. Chen, Microcontroller-based B-spline neural position control for voice coil motors, IEEE Trans. Ind. Electron. 62 (2015) 56445654.
[33] K.J. Astrom, B. Wittenmark, Adaptive Control, Addision-Wesley, New York, 1995.
[34] J.R. Leigh, Applied Digital Control: Theory, Design, and Implementation, second ed., Dover Publications, New York, 2006.